Machine learning to optimize climate projection over China with multi-model ensemble simulations

Tong Li, Zhihong Jiang, Hervé Le Treut, Laurent Li, Lilong Zhao, Lingling Ge

Environmental Research Letters · 12 August 2021 · Nanjing University of Information Science & Technology (NUIST)

In one sentence. Random forest combines CMIP6 multi-model ensembles over China, projecting less warming and wetter conditions with better spatial skill than the arithmetic mean.

Abstract

Abstract The multi-model ensemble approach is generally considered as the best way to explore the advantage and to avoid the weakness of each individual model, and ultimately to achieve the best climate projection. But the design of an optimal strategy and its practical implementation still constitutes a challenge. Here we use the random forest (RF) algorithm (from the category of machine learning) to explore the information offered by the multi-model ensemble simulations within the Coupled Model Intercomparison Project Phase 6. Our objective is to achieve a more reliable climate projection (mean climate and extremes) over China. RF is furthermore compared to two other ensemble-processing strategies of different nature, one is the basic arithmetic mean (AM), and another is the linear regression across the ensemble members. Our results indicate that RF effectively enhances the capability in capturing spatial climate characteristics. Regions with complex topography, such as the Tibetan Plateau and its periphery, show the most significant improvements. RF projects less future warming but enhanced wet conditions across China. It also produces larger spatial variability and more small-scale features. The most obvious increase of precipitation is in the northern part and the periphery of the Tibetan Plateau. The projected changes in RF for strong precipitation are almost twice higher than in AM, while in the northwestern area, weaker increases of precipitation are projected by RF, which indicates larger spatial inhomogeneity of its projection.

Labels

Labels and summary are generated by Claude and reviewed weekly. Wrong? Edit this entry.